How Duke University Is Working to Make AI More Sustainable

— and to build a more sustainable world

drone launch by person to assist climate research
Photo courtesy of David Johnston, PhD

Artificial intelligence is transforming research, accelerating discovery and helping scientists tackle increasingly complex challenges. But as the data centers that house the AI models become larger and more powerful, they can consume vast amounts of electricity and water, raising critical questions: Can AI become more sustainable? Could it actually help address climate change?

At Duke University, researchers are investigating both of those challenges. Across disciplines, faculty are developing more energy-efficient AI technologies while using AI to improve climate modeling, strengthen weather forecasting, protect oceans and modernize energy systems. The university is expanding its AI infrastructure with a small GPU center expected to open in 2027, which will be designed to minimize power and water consumption and carbon emissions through energy-efficient practices.

 “Climate change is one of the defining challenges of our time, and AI can be an important part of the solution, but only if we develop it responsibly. Duke's commitment is to lead in both directions: advancing AI that helps society understand and respond to a changing planet while pioneering the energy-efficient, ethical, and equitable technologies needed to make AI itself more sustainable," said Toddi Steelman, Duke’s vice president and vice provost for climate and sustainability.

AI is already improving predictions of global temperatures, ocean conditions and extreme weather events while helping policymakers evaluate climate adaptation strategies and optimize energy systems. But its rapid growth is also increasing demand for electricity.

“AI and data centers have pushed us from a world of flat-load growth in electricity to one of rising demand,” said Brian Murray, director of Duke's Nicholas Institute for Energy, Environment & Sustainability. “Large data center power demands are expected to rise about 130% by the end of this decade, and about one-half of the load growth is due to AI.”

A first-of-its-kind analysis released by Nicholas Institute researchers found that taking advantage of load flexibility could enable the U.S. power system to more quickly absorb the new demand while mitigating the immediate need for costly expansion of grid capacity.

As a Duke Ph.D. student, Tyler Norris worked with Duke faculty on the research, which has had policy and industry impact. The research has helped catalyze industry and policy discussions about large-load flexibility, and led to policy recommendations.

“We’ve never planned the system this way before,” Norris said in a recent Duke feature. “[Load] flexibility could help maintain reliability, control costs and reduce emissions.”

Efficient AI Efforts

Duke researchers are also developing new ways to think about how AI can be built more efficiently.

Three people looking down at a device
Helen Li is among several Pratt faculty working on new hardware approaches to reduce the water and energy consumption of AI. Photo courtesy of Pratt.
A woman posing in a computer lab
Pratt Professor Tania Roy's team is working on semiconductor devices capable of performing AI tasks without the need of a data center. Photo courtesy of Pratt.

Electrical and Computer Engineering Professor Tania Roy is working on innovative approaches to reduce the water and energy consumption of AI hardware. Roy’s laboratory is creating neuromorphic—or brain-inspired—semiconductor devices capable of performing AI tasks on their own, without the need of a data center.

“The transistor is the basic device that has revolutionized everything,” Roy said. “Our computers, mobile phones—everything has them. We want to change that basic building block for AI, so that it can do AI more efficiently.”

Her team is developing tiny experimental circuits that mimic the efficiency of the human brain. Instead of sending every AI request to an energy-intensive server farm, future smart devices based on chips, like the prototypes they are building, could process information locally. Roy describes a future where voice assistants, autonomous vehicles and smart cameras perform sophisticated AI tasks on the device itself to reduce energy load.

Also in Pratt, Helen Li, Marie Foote Reel E’46 Distinguished Professor, and Yiran Chen, John Cocke Distinguished Professor, are working on different hardware approaches with the same goal: to reduce the water and energy consumption of AI. Their award-winning work on magnetic random access memory is contributing to future and more efficient AI.

Two people sitting at a table surrounded by computer equipment
Yiran Chen and colleagues are developing robotic AI processing hardware with the goal to be more energy efficient.

Chen and Roy are among researchers recently awarded a project funded by the Department of Energy to develop robotic AI processing hardware that is significantly faster and more efficient on local hardware—with the intent to reduce the amount of energy data sent to and from data centers.

Understanding Climate Extremes

While engineers work to reduce AI’s energy demands, Duke climate scientists are also using the technology to better understand Earth’s changing climate.

In the Nicholas School of the Environment, Professor Shineng Hu applies machine learning and deep learning to improve forecasts of climate extremes such as El Niño events and climate-related societal impacts, including malaria outbreaks in South America.

His Climate Dynamics Lab combines observations, theoretical research and sophisticated climate models to study how oceans and the atmosphere interact across timescales. This work ranges from daily weather scales to millions of years of Earth's history. The topics Hu Lab investigate include global warming patterns, El Niño, marine heatwaves, atmospheric rivers, Atlantic overturning circulation, and more.

AI helps his team recognize complex climate patterns hidden within enormous datasets, improving scientists' ability to forecast future conditions and understand how climate change influences extreme weather.

AI is also changing how researchers study the natural world—and informing how to respond to changes.

Observing Wildlife and Coastal Ecosystems

As climate change rapidly reshapes marine habitats, Duke’s Marine Robotics and Remote Sensing Lab, led by director David Johnston, is combining drones, satellites and artificial intelligence to help scientists observe wildlife and coastal ecosystems at the speed and scale required for effective conservation.

“In the past, scientists didn’t have enough data. Now the script is flipping, and we are often data rich but without enough time to go through all the data. AI can help us analyze our data in 20% of the time, and we can rapidly see the results.”

David Johnston, Duke Marine Robotics and Remote Sensing Lab

The lab uses AI to turn imagery collected by drones and satellites into practical information for marine conservation. Its work addresses a basic problem facing climate and biodiversity science: Researchers can now collect enormous quantities of high-resolution environmental imagery, but manually finding, counting, identifying and measuring animals or habitats within those images can take months or years.

“In the past, scientists didn’t have enough data. Now the script is flipping, and we are often data rich but without enough time to go through all the data. AI can help us analyze our data in 20% of the time, and we can rapidly see the results,” he said.

Through studies of whales, harbor seals, sea turtles and seabirds—as well as oyster reefs, wetlands and coastal forests— the lab has developed systems that can document a range of characteristics. These systems can detect animals, estimate population size, identify species, measure body dimensions, map habitat and document environmental change. The aim is to reduce analytical bottlenecks so that researchers and managers can monitor vulnerable species more frequently across larger areas and in places that are dangerous or difficult to reach.

Duke’s Marine Robotics and Remote Sensing Lab, led by director David Johnston, is combining drones, satellites and artificial intelligence for conservation. Photos courtesy of David Johnston.

With scientists involved at every step, AI models detect and count animals, identify species, measure bodies, map habitats and reconstruct change through time.

The lab has developed an integrated workflow: Remote sensing expands observation. Drones collect detailed information safely and repeatedly, while satellites extend coverage across larger and more remote regions. AI converts imagery into ecological measurements. Models detect and count animals, identify species, measure bodies, map habitats and reconstruct change through time. Scientists are involved at every step to label training data, verify detections, evaluate uncertainty and interpret results in their ecological and management context.

Tracking Environs, from Glaciers to Disease Outbreaks

The combination of drones and AI is now being used to continue long-term monitoring of harbor seals in Glacier Bay National Park and Preserve in Alaska as they adapt to receding glaciers.

“Harbor seals are dependent on glacial ice, but the glaciers are receding due to global warming,” Johnston said. “Using autonomous drones programmed to photograph their habitat, we generate thousands of images. Turning that into knowledge requires analysis. With explainable AI, the computer is looking at the same thing we are looking at and we can validate the data in the images.”

Marta Zaniolo, civil and environmental engineering professor, pulls AI concepts into complex models of local and regional water use. Working on water resources management, Zaniolo combines hydrology and climatology with machine learning and data mining to make better decisions about water use.

Zaniolo recently worked on a dam construction project between Ethiopia and Kenya. Dealing with multiple factors, she trained a computer model to use control and optimization techniques typically employed in robotics and apply them to issues in water resources. She has also worked with U.S. cities to use the techniques to help with resource planning.

Duke researchers have also developed an AI tool to predict when and where malaria outbreaks will occur— months in advance— to help reduce a global health threat.

Several people in an outside tent.
Researchers meet as part of a study tracking mosquitoes in a heavily forested region of Panama, where malaria cases have risen 15-fold in the past four years. Photo courtesy of William Pan.

Created by biostatistician William Pan of the Nicholas School of the Environment and the Duke Global Health Institute and colleagues, the system analyzes variables including temperature, precipitation and satellite imagery to forecast where and when malaria is likely to spread. These insights are helping countries to deploy targeted interventions such as diagnostic tests and medications.

Supported by NASA and the National Institute of Allergy and Infectious Diseases, the system has been validated for malaria forecasting in five countries: Peru, Ecuador, Brazil, Panama and Honduras. The system has also been tested in Colombia, and Pan and his team is evaluating expansion to other threats like dengue and leptospirosis. 

Duke’s AI strategy extends beyond disciplinary boundaries— and is focused on solutions for people and communities.

An Ethical AI Infrastructure

The Deep Tech at Duke Initiative, led by Interim Director of the Duke Initiative for Science & Society David Hoffman, brings together researchers and educators across disciplines to advance science and policy with a focus on transparency, accountability and ethical governance.

Hosted by program lead Merritt Cahoon, Deep Tech held a recent event called The Infrastructure Imperative: AI’s Physical Foundation. The event aimed to examine the systems required to build, finance, power and scale the next generation of AI infrastructure: software, but also data centers, power generation, cooling, water, permitting, construction timelines and capital investment.

The initiative also brought together a working group of leaders from the public and private sectors to assess emerging challenges in AI infrastructure development and deployment, share research insights from Duke faculty, and strengthen cross-sector collaboration among academia, industry and government.

Steelman emphasized the value of universities in this technological transformation:  "Universities have a unique responsibility in this moment. We bring together engineers, climate scientists, social scientists, policymakers and ethicists to ask not only what AI can do, but what it should do. That’s how we'll develop technologies that benefit society while protecting the planet."